refactor(moe): extract hybrid decode orchestration

[zihaomu]

Summary

Extract the shared Hybrid MoE decode schedule from OffloadMoELayer into
HybridDecodeExecutor. This is now a standalone, single-commit refactor on
main (9b585b7), not a stack containing #378.

• Preserve the current LRU/fetch policy, in-place route rewriting, optional
  statistics, CPU submit/sync order, and FREETOKEN_HYBRID_OVERLAP=0 behavior.
• Keep quantization dispatch and supports_inactive_slots handling in the layer:
  capable kernels receive -1; other kernels get zero-weighted slot 0.
• Keep cache ownership acyclic and retain the executor across cache rebuilds.
• Add routing/order/numerical/lifetime tests and make an existing offload-config
  test use Triton rather than an unrelated CUDA-only FlashInfer selection.

No changes to CPU kernels, CUDA/HIP synchronization, cache replacement policy,
or expert math. The short design note describes these boundaries.

Related: #350 and #378. This PR does not fix ROCm CPU/Hybrid graph safety;
#378 remains a separate change. Neither its implementation nor its validation
claims are included here.

Validation

One AMD Radeon AI PRO R9700 (gfx1201, 32 GiB); dual EPYC 9334; Ubuntu 24.04;
host kernel/driver 7.0.0-31-generic; PyTorch 2.11.0+rocm7.14.0,
HIP 7.14.60850, Triton 3.7.1. Synthetic expert banks; no model checkpoint.
Only one GPU was exposed to the container.

python setup.py build_ext --inplace
PYTHONPATH=python HIP_VISIBLE_DEVICES=0 CUDA_VISIBLE_DEVICES=0 \
  OMP_NUM_THREADS=4 MKL_NUM_THREADS=4 taskset -c 32-39 python -m pytest -q \
  tests/moe/test_hybrid_decode_executor.py \
  tests/moe/test_hybrid_fetch.py tests/moe/test_offload.py \
  tests/moe/test_cpu_moe.py tests/moe/test_cpu_moe_q4_0.py \
  tests/engine/test_moe_cpu_layers.py \
  tests/engine/test_attention_backend_arch.py \
  tests/engine/test_attention_backend_matrix.py \
  tests/models/test_quant_config.py \
  -m "not slow" -k "not cuda_graph"

181 passed, 42 skipped, 5 deselected. The new file contributes 21 CPU
contract cases and 12 real BF16 Hybrid CPU/GPU cases. The existing CPU-executor
tests additionally cover BF16, NVFP4, MXFP4, DS-FP4 and Q4_0 eager computation;
this is not five-format Hybrid end-to-end coverage.

A supplemental local A/B harness compared the saved main layer implementation
against this branch using identical cache/kernels/CPU executor: 36 combinations
x 12 steps, bitwise-equal outputs and rewritten routing, identical cache state
and statistics. Batch sizes 1/4/8, fetch caps 0/1/8, overlap on/off and
router-weight-on-input on/off were exercised.

Local single-layer eager microbenchmark (not a shipped benchmark or model-serving
result): BF16, 3 layers, 8 experts, H=1024, I=512, top-k=2, fetch cap=1,
4 CPU workers, CPUs 32-39, warmup then 10 alternating A/B rounds of 128 calls.

Batch	Overlap	main (us/layer)	refactor (us/layer)	Change
1	off	313.01	315.75	+0.88%
1	on	325.73	328.44	+0.83%
4	off	434.34	434.37	+0.01%
4	on	370.60	370.98	+0.10%

Limits

• No NVIDIA hardware, full-model serving/throughput, or CPU/Hybrid graph replay
  validation in this revision. The four graph-replay tests and one slow test
  were explicitly deselected above; graph support is inherited from main.
• Initial unrestricted-affinity diagnostics intermittently waited a long time
  in HIP synchronization, including while executing the saved main path;
  two runs were interrupted. The completed A/B and pinned-core final suite and
  benchmark passed. This refactor does not claim to diagnose or fix those waits.

Ready for review as a focused orchestration refactor, not as a graph-safety fix
or a claim of zero overhead/full-model performance parity.

[zihaomu]

Rebased this proposal onto the refreshed CPU/Hybrid ROCm graph-safety PR (#378, head 94f7724) and ported the extraction to the current upstream MoE architecture.

The conflict in layers/moe.py was resolved semantically: the current moe_strategy and quant-method dispatch remain intact, while Hybrid route splitting / CPU-GPU overlap / result merging move into HybridDecodeExecutor. OffloadMoELayer now supplies only the narrow cached-GPU-expert callback. No legacy MoE implementation was restored.

Validation:

• Hybrid executor contracts + ROCm graph-safety tests: 21 passed
• adjacent Hybrid fetch tests: 3 passed
• CPU MoE suites: 30 GPU-dependent cases collected/skipped in the no-GPU container
• Python compile and git diff --check pass

New head: d5fa33231edd2328370fe7a7968b00701147c082.